Photovoltaic load prediction method and system based on multi-source data deep learning

Through multi-source data set training multiple deep learning models and weighted fusion prediction results, the deviation problem caused by a single data source in photovoltaic load prediction is solved, the accuracy and reliability of prediction is improved, and real-time data update and incremental training is carried out. model, enhances the dynamic adaptability of the predicted results.

CN119009969AInactive Publication Date: 2024-11-22JIANGSU JOYRUN INFORMATION TECH CO LTD
View PDF 0 Cites 9 Cited by

Patent Information

Application Number
CN202411063743.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing photovoltaic load prediction methods rely on a single data source and ignore interference factors such as power generation power, power generation efficiency, transmission loss and real-time weather data, resulting in a large deviation from the prediction results and the actual load cannot meet the needs of high-precision prediction.

Method used

Using deep learning methods based on multi-source data, integrate historical interference data to build multi-source data sets, train multiple different deep learning models, and improve the accuracy and reliability of predictions by weighted fusion of their prediction results, and update them through real-time data. and incremental training of the model to enhance dynamic adaptability.

Benefits of technology

By integrating multi-source data for deep learning model training, we can more comprehensively capture the key factors affecting photovoltaic load, improve the accuracy and reliability of photovoltaic load prediction, and enhance the dynamic adaptability and long-term stability of the prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119009969A_ABST
    Figure CN119009969A_ABST
Patent Text Reader

Abstract

The invention discloses a photovoltaic load prediction method and system based on multi-source data deep learning, and relates to the technical field of new energy photovoltaic load prediction, in the method, the photovoltaic load prediction system constructs a multi-source data set by integrating multiple types of interference data influencing a photovoltaic load prediction result. On the basis, a plurality of different deep learning models are trained by using a multi-source data set, respective advantages of different models are exerted, and a complex photovoltaic load can be predicted from different emphasis. And the prediction results of the multiple models are subjected to weighted fusion according to the accuracy, so that the robustness and accuracy of prediction are further improved. By implementing the method, the defects of single data source and large prediction deviation in the photovoltaic load prediction process in related technologies are overcome, deep learning model training is performed by integrating multi-source data, and key factors influencing the photovoltaic load can be captured more comprehensively, so that the accuracy and reliability of photovoltaic load prediction are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of new energy photovoltaic load prediction, and in particular to a photovoltaic load prediction method and system based on multi-source data deep learning. Background Art

[0002] Photovoltaic load forecasting refers to the activity of predicting the future power generation of photovoltaic power stations using information such as weather forecasts and historical power generation data. Through photovoltaic load forecasting, power generation plans can be better arranged, supply and demand can be balanced, power fluctuations can be reduced, stable data support can be provided for power grid operation, and energy utilization can be improved. At the same time, it can provide reference for the investment and construction of photovoltaic power stations, assist power market operations, and promote emission reduction.

[0003] With the rapid development of new energy photovoltaic technology, photovoltaic load forecasting has become a key link to ensure the stable operation of the power system. At present, a variety of data-driven load forecasting methods have been proposed in related technologies, which mainly rely on historical photovoltaic load data for model training, and then input the currently detected environmental data and power generation power into the trained model to obtain the final photovoltaic load forecast result.

[0004] However, the photovoltaic load prediction method in the related technology mainly relies on historical photovoltaic load data as its data source. The data source is single and other interference factors that affect the photovoltaic load prediction results, such as power generation, power generation efficiency, transmission loss and real-time weather data, are ignored. As a result, there is a large deviation between the prediction results and the actual load, which cannot meet the needs of high-precision prediction. Summary of the invention

[0005] The present application provides a photovoltaic load prediction method and system based on multi-source data deep learning, which is used to integrate multi-source data that affect the photovoltaic load prediction results, and obtain the photovoltaic load prediction results by parallel processing of multiple different deep learning models.

[0006] In a first aspect, the present application provides a photovoltaic load forecasting method based on multi-source data deep learning, which is applied to a photovoltaic load forecasting system. The method includes: Acquire historical interference data and construct a multi-source data set, where the historical interference data includes power generation data, power generation efficiency data, transmission loss data, historical weather data, and historical load data of the photovoltaic power generation system; Use multi-source data sets to train multiple different deep learning models respectively, and obtain multiple different target deep learning models; Get the weighted value corresponding to each target deep learning model; The multi-source data acquired in real time are input into different target deep learning models respectively, and the photovoltaic load prediction results corresponding to each target deep learning model are output. The multi-source data corresponds to the historical interference data one by one. The final photovoltaic load prediction result is calculated based on the photovoltaic load prediction result output by each target deep learning model and the corresponding weighted value.

[0007] Through the above embodiments, the photovoltaic load forecasting system constructs a multi-source data set by integrating various types of interference data that affect the photovoltaic load forecasting results. On this basis, multiple different deep learning models are trained with the multi-source data set to give full play to the respective advantages of different models, so that complex photovoltaic loads can be predicted from different perspectives. The prediction results of multiple models are then weighted and fused according to their accuracy, further improving the robustness and accuracy of the prediction. This method solves the defects of the related technology in the process of photovoltaic load forecasting, such as a single data source and large prediction deviation. By integrating multi-source data for deep learning model training, the key factors affecting photovoltaic load can be captured more comprehensively, thereby improving the accuracy and reliability of photovoltaic load forecasting.

[0008] In some embodiments, after the step of using multiple source data sets to train multiple different deep learning models respectively to obtain multiple different target deep learning models, the method further includes: Acquire new multi-source data in real time; Update multi-source datasets based on new multi-source data; Regularly use new data from multi-source datasets to incrementally train the target deep learning model.

[0009] Through the above embodiments, the photovoltaic load forecasting system continuously acquires new data generated during the real-time operation of the photovoltaic power generation system, and regularly performs incremental learning on the trained deep learning model, so that the model can continuously track the changing trend of photovoltaic load characteristics, adaptively optimize and update the forecasting model, and enhance the dynamic adaptability and long-term stability of the photovoltaic load forecasting results.

[0010] In some embodiments, the step of obtaining the weighted value corresponding to each target deep learning model specifically includes: Get the accuracy of each target deep learning model on the validation set; The weighted value corresponding to each target deep learning model is determined according to the ratio between the accuracies.

[0011] Through the above embodiment, the photovoltaic load forecasting system evaluates the contribution of each model to the final forecast result based on the forecast accuracy of the model on the validation set, and the model with higher accuracy is given a greater weight. This method can make fuller use of the forecast results of models with good performance and weaken the influence of poor models, so that the forecast results of multi-model fusion are as close to the true value as possible.

[0012] In some embodiments, the steps of using multiple source data sets to train multiple different deep learning models respectively to obtain multiple different target deep learning models specifically include: According to the training requirements of each deep learning model, the multi-source data set is preprocessed to obtain multiple feature training sets, which correspond one-to-one to the deep learning model; The deep learning model is trained using the feature training set to obtain multiple different target deep learning models.

[0013] Through the above embodiments, different deep learning models have different requirements for the format and quality of training data. In order to give full play to the modeling capabilities of each model, the photovoltaic load forecasting system performs customized preprocessing on the original multi-source data set according to the characteristics of each deep learning model, and extracts feature training sets that can match each model interface respectively. It not only takes into account the needs of multi-model integration strategies for data diversity, but also takes into account the personalized requirements of different models for data quality. While improving the prediction performance of a single model, it also lays the foundation for subsequent model fusion.

[0014] In some embodiments, the steps of using multiple source data sets to train multiple different deep learning models respectively to obtain multiple different target deep learning models specifically include: Divide the multi-source data set according to the processing data type preset by each deep learning model to obtain a dedicated training set corresponding to each deep learning model, where the dedicated training set is a subset of the multi-source data set; The deep learning model is trained using a dedicated training set to obtain multiple different target deep learning models.

[0015] Through the above embodiments, in order to enable each model to play its strengths, the photovoltaic load forecasting system selects the data subset that each model is best at processing from the multi-source data set as its exclusive training set. Each trained model can make reliable predictions in its respective areas of expertise, thereby improving the accuracy and reliability of photovoltaic load forecasting.

[0016] In some embodiments, after the step of calculating the final photovoltaic load prediction result according to the photovoltaic load prediction result output by each target deep learning model and the corresponding weighted value, the method further includes: Get actual measurements of PV load; Generate an error curve of photovoltaic load prediction based on the difference between the actual measurement value and the final photovoltaic load prediction result; The final photovoltaic load prediction results are evaluated based on the fluctuation of the error curve.

[0017] Through the above embodiments, the photovoltaic load prediction system introduces real photovoltaic load measurement values ​​as evaluation criteria, quantifies the deviation between the multi-model fusion prediction results and objective facts, and visually displays the dynamic change trend of the prediction error over time through the error curve, so as to facilitate users to evaluate the accuracy of the final photovoltaic load prediction results.

[0018] In some embodiments, the step of evaluating the final photovoltaic load prediction result according to the fluctuation condition of the error curve specifically includes: Count the number of final photovoltaic load forecast results whose difference exceeds a preset range threshold within a preset time period; If it is detected that the number of final photovoltaic load prediction results exceeds a preset number threshold, it is determined that the final photovoltaic load prediction result is abnormal.

[0019] Through the above embodiment, when the photovoltaic load forecasting system detects that the forecast error remains at a low level for a long time and the fluctuation is small, it determines that the final photovoltaic load forecast result generated by the forecast model (target deep learning model) is normal; on the contrary, if the error is large or fluctuates, it determines that the final photovoltaic load forecast result is abnormal. Automated monitoring of the forecast results of the target deep learning model is achieved, which is convenient for users to promptly discover problems with the model and deal with them.

[0020] In a second aspect, the present application provides a photovoltaic load forecasting system, the photovoltaic load forecasting system comprising: one or more processors and a memory; The memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions. The one or more processors call the computer instructions so that the photovoltaic load forecasting system can implement a photovoltaic load forecasting method based on deep learning of multi-source data provided in the above embodiment, which will not be repeated here.

[0021] In a third aspect, the present application provides a computer-readable storage medium comprising instructions. When the instructions are executed on a photovoltaic load forecasting system, the photovoltaic load forecasting system can implement a photovoltaic load forecasting method based on deep learning of multi-source data provided in the above-mentioned embodiment, which will not be repeated here.

[0022] In a fourth aspect, the present application provides a computer program product. When the computer program product runs on a photovoltaic load forecasting system, the photovoltaic load forecasting system can implement a photovoltaic load forecasting method based on deep learning of multi-source data provided in the above embodiment, which will not be repeated here.

[0023] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. It solves the defects of single data source and large prediction deviation in the process of photovoltaic load forecasting in related technologies. By integrating multi-source data for deep learning model training, it can more comprehensively capture the key factors affecting photovoltaic load, thereby improving the accuracy and reliability of photovoltaic load forecasting.

[0024] 2. The idea of ​​ensemble learning is introduced, multiple deep learning models are trained using multi-source data, and the prediction results of multiple deep learning models are integrated to further improve the accuracy and stability of photovoltaic load prediction.

[0025] 3. The photovoltaic load forecasting system continuously acquires new data generated during the real-time operation of the photovoltaic power generation system, and regularly performs incremental learning on the trained deep learning model, so that the model can continuously track the changing trend of photovoltaic load characteristics, adaptively optimize and update the prediction model, and enhance the dynamic adaptability and long-term stability of the photovoltaic load forecasting results. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of a photovoltaic load prediction method based on multi-source data deep learning in an embodiment of the present application; Figure 2 It is another flow chart of a photovoltaic load prediction method based on multi-source data deep learning in an embodiment of the present application; Figure 3 It is a schematic diagram of a process of evaluating the final photovoltaic load prediction result by the photovoltaic load prediction system in the embodiment of the present application; Figure 4 It is a schematic diagram of the structure of a physical device of the photovoltaic load prediction system in the embodiment of the present application. DETAILED DESCRIPTION

[0027] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to be used as limitations to the present application. As used in the specification and appended claims of the present application, the singular expressions "one", "a kind of", "said", "above", "the" and "this" are intended to also include plural expressions, unless there is a clear indication to the contrary in the context. It should also be understood that the term "and / or" used in the present application refers to any or all possible combinations comprising one or more listed items.

[0028] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as suggesting or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features, and in the description of the embodiments of the present application, unless otherwise specified, "plurality" means two or more.

[0029] For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

[0030] Photovoltaic load forecasting is a key link in the stable operation of the power system. With the rapid development of new energy photovoltaic technology, higher requirements are placed on its prediction accuracy. Take a photovoltaic power station as an example. The installed capacity of the power station is 50MW, and the annual power generation is 80 million kWh. It is an important part of the local power grid. The grid dispatching department needs to accurately grasp the power generation of the photovoltaic power station in the future (such as a week) in order to reasonably arrange the power dispatching plan and balance the supply and demand of the power grid. However, the factors affecting photovoltaic power generation are complex, including meteorological conditions such as light intensity, temperature, and humidity, as well as the power generation efficiency of photovoltaic modules and the loss of transmission lines. How to mine the implicit influence rules from massive historical data and establish a reliable photovoltaic load forecasting model has become a major challenge facing the power system.

[0031] In response to the above-mentioned needs of photovoltaic load forecasting, traditional methods usually train forecasting models based on a single data source (such as historical load data). Taking the above-mentioned 50MW photovoltaic power station as an example, forecasters extracted load sequences as samples from the hourly load records of the power station in the past year, and used machine learning algorithms such as support vector machines for training to obtain a load forecasting model. In actual forecasting, the load data of the recent period is input into the model to obtain the load forecast value for the future period. Although this method is simple to implement, it ignores other influencing factors besides load data, such as weather changes, equipment status, etc., resulting in insufficient generalization performance of the model and difficulty in making accurate predictions.

[0032] To solve the above problems, the present application provides a photovoltaic load forecasting method and system based on deep learning of multi-source data. Specifically, the above-mentioned 50MW photovoltaic power station is used as an example for introduction. The photovoltaic load forecasting system first collects multi-source data from various business departments. In addition to conventional load data, it also includes hourly irradiance, temperature and other meteorological data, real-time power generation and conversion efficiency data of photovoltaic components, and interference data such as power loss of transmission lines. These interference data are unified in format to construct a multi-source data set. Then, the photovoltaic load forecasting system uses the data set to train multiple different deep learning models, such as long short-term memory network (LSTM), gated recurrent unit network (GRU), etc., to make full use of the advantages of each model. In actual prediction, the photovoltaic load forecasting system inputs multi-source data into each trained model to obtain multiple prediction results, and then calculates weights according to the prediction accuracy of each model on the validation set, and weights and fuses multiple results to obtain the final photovoltaic load prediction value.

[0033] For ease of understanding, the following describes the process of the method provided by this implementation in combination with the above scenario. Figure 1 , which is a flow chart of a photovoltaic load prediction method based on deep learning of multi-source data in an embodiment of the present application.

[0034] S101. Obtain historical interference data and construct a multi-source data set.

[0035] The photovoltaic load forecasting system collects various historical interference data that affect photovoltaic power generation load from multiple business departments and monitoring equipment. These data include but are not limited to the power generation data, power generation efficiency data, transmission loss data, historical weather data and historical load data of the photovoltaic power generation system. Furthermore, after pre-processing operations such as data cleaning, data alignment, and data normalization on the above historical interference data, the photovoltaic load forecasting system constructs a multi-source data set according to the preset data structure and format requirements, which is used as the basic data for the subsequent training of multiple different deep learning models.

[0036] It should be noted that the data structure and format requirements of the multi-source dataset are related to the deep learning model selected subsequently, and can be adjusted according to actual needs and are not limited here.

[0037] S102: Use multi-source data sets to train multiple different deep learning models respectively to obtain multiple different target deep learning models.

[0038] The photovoltaic load forecasting system uses the multi-source data set constructed in step S101 to train multiple different types of deep learning models selected by the user, in order to learn the implicit laws affecting the photovoltaic load from different angles and levels. Compared with the traditional single model, this application adopts a multi-model integration strategy, integrating the learning capabilities of multiple models, and can better fit the complex photovoltaic load fluctuation law.

[0039] Specifically, users can select deep learning models for training according to actual needs, such as LSTM, GRU, and CNN. The number of deep learning models involved in training is at least two or more, where different deep learning models selected have their specific advantages, such as LSTM is good at capturing long-term dependencies in time series data, CNN is good at extracting spatial features, DBN can effectively perform unsupervised feature learning, etc. Users can determine the number and type of deep learning models involved in training according to actual needs, and there is no limitation here.

[0040] To facilitate understanding, the following takes a deep learning model based on the LSTM algorithm as an example to introduce the process of training the model using multi-source datasets.

[0041] First, the photovoltaic load forecasting system preprocesses the multi-source data set and integrates the data of the multi-source data set into a time series form suitable for the LSTM model input. Specifically, the multi-source data set is first time-aligned to ensure that all data sources (power generation, power generation efficiency, transmission loss, weather data, and historical load data) are aligned in the time dimension. Usually a fixed time interval is selected, such as 15 minutes or 1 hour. Then the time features in the data source are extracted, normalized and serialized to obtain the preprocessed multi-source data set.

[0042] Furthermore, the photovoltaic load forecasting system uses the above preprocessed multi-source data set to train the LSTM model, and evaluates the performance of the LSTM model through indicators such as the root mean square error (RMSE) on the test set. The LSTM model with the largest evaluation index and meeting the preset threshold is selected for subsequent photovoltaic load forecasting.

[0043] It should be noted that the training process of models such as GRU and CNN is similar to the training process of the above-mentioned LSTM model and will not be repeated here.

[0044] S103. Obtain a weighted value corresponding to each target deep learning model.

[0045] After obtaining multiple different target deep learning models, the photovoltaic load forecasting system needs to assign a weight to each trained target deep learning model to reflect the importance of the model in the final forecast result. To this end, the photovoltaic load forecasting system uses a pre-prepared validation data set to evaluate the prediction performance of each target deep learning model. The validation data set is independent of the model's training data set, but has a similar data distribution.

[0046] Furthermore, the photovoltaic load forecasting system calculates the prediction accuracy of each model on the validation data set. For example, an accuracy calculation formula is: accuracy = 1 / (1 + error index). The evaluation index can be the mean absolute error (MAE), the root mean square error (RMSE), or the mean absolute percentage error (MAPE), etc., which are not limited here. In addition, the above accuracy calculation formula is not unique and can be adjusted according to actual needs, which is not limited here.

[0047] Furthermore, the photovoltaic load forecasting system determines the weight of each model according to the accuracy score calculated by the above accuracy calculation formula. For example, the accuracy score can be converted into a probability distribution using a softmax function to ensure that the sum of all weights is 1. The ratio of the accuracy score can also be directly used as the weight, that is, the weight of each model is equal to its accuracy divided by the sum of the accuracies of all models. This is not limited here.

[0048] For ease of understanding, let's take an example. For example, suppose the photovoltaic load forecasting system trains three models: LSTM, CNN, and DBN. On the validation set, their mean absolute percentage errors (MAPE) are 5%, 6%, and 7%, respectively. The photovoltaic load forecasting system obtains their accuracies of 0.952, 0.943, and 0.935, respectively, based on the above accuracy calculation formula. Then, these accuracies are normalized to obtain the final weights of 0.337, 0.333, and 0.330, respectively.

[0049] S104: Input the multi-source data acquired in real time into different target deep learning models respectively, and output the photovoltaic load prediction result corresponding to each target deep learning model.

[0050] Specifically, the photovoltaic load forecasting system first needs to obtain real-time multi-source data corresponding to historical interference data, including current power generation, power generation efficiency, transmission loss, real-time weather data, and the latest load data. These real-time multi-source data can be obtained through various sensors, monitoring equipment, weather stations and other real-time collection systems, which will not be described here.

[0051] Next, the photovoltaic load forecasting system preprocesses these real-time multi-source data to make their format and scale consistent with the training data. The processed real-time multi-source data is then input into each previously trained target deep learning model. Each model will generate a photovoltaic load forecast result based on the input real-time data and the learned patterns and features.

[0052] The forecast results output by each model may be the expected photovoltaic load value at a certain point in the future (such as 15 minutes later, 1 hour later), or the load curve within a period of time (such as the next 24 hours). The photovoltaic load forecasting system will save the forecast results of each model to prepare for the next integrated forecast.

[0053] S105. Calculate the final photovoltaic load prediction result based on the photovoltaic load prediction result output by each target deep learning model and the corresponding weighted value.

[0054] Specifically, the photovoltaic load forecasting system first retrieves the photovoltaic load forecasting results output by each target deep learning model and the corresponding weighted values ​​determined in step S103. Then, a weighted average method is used to calculate the final forecasting result. Specifically, for each forecast time point, the photovoltaic load forecasting system multiplies the forecast value of each model by its corresponding weight, and then adds these weighted forecast values ​​to obtain the final forecast value at that time point.

[0055] For example, suppose the photovoltaic load forecasting system uses three models (LSTM, CNN and DBN), and their photovoltaic load forecast values ​​for the next hour are 80kW, 85kW and 82kW, respectively, with corresponding weights of 0.4, 0.35 and 0.25, respectively. Then, the final forecast value calculated by the system will be: 80×0.4+85×0.35+82×0.25=82.15kW.

[0056] Through the above embodiments, the photovoltaic load forecasting system constructs a multi-source data set by integrating various types of interference data that affect the photovoltaic load forecasting results. On this basis, the multi-source data set is used to train multiple different deep learning models, giving play to the respective advantages of different models, and can make predictions for complex photovoltaic loads from different perspectives. The prediction results of multiple models are then weighted and fused according to their accuracy, further improving the robustness and accuracy of the prediction. This method solves the defects of the single data source and large prediction deviation in the process of photovoltaic load forecasting in related technologies. By integrating multi-source data for deep learning model training, it can more comprehensively capture the key factors affecting photovoltaic loads, thereby improving the accuracy and reliability of photovoltaic load forecasting.

[0057] The following is a more detailed description of the process of the method provided by this embodiment. Figure 2As shown, it is another flow chart of a photovoltaic load prediction method based on multi-source data deep learning in an embodiment of the present application.

[0058] S201 . Update the multi-source data set according to the acquired new multi-source data.

[0059] The photovoltaic load forecasting system receives the latest multi-source data uploaded by multiple business departments and monitoring equipment, and the multi-source data type corresponds to the historical interference data one by one. For example, if the historical interference data in the multi-source data set includes the power generation data, power generation efficiency data, transmission loss data, historical weather data and historical load data of the photovoltaic power generation system, the new multi-source data also includes the power generation data, power generation efficiency data, transmission loss data, weather data within the preset time and historical load data of the photovoltaic power generation system that is currently detected.

[0060] On this basis, the multi-source data set is updated based on the newly acquired multi-source data to improve the timeliness of the interference data in the multi-source data set.

[0061] S202. Regularly use the newly added data in the multi-source dataset to perform incremental training on the target deep learning model.

[0062] First, the user can set an incremental training cycle on the photovoltaic load forecasting system. This cycle can be daily, weekly or monthly. The specific frequency can be determined based on the frequency of data updates and the availability of computing resources, which is not limited here. Then, at the beginning of each training cycle, the photovoltaic load forecasting system extracts the new data since the last training from the updated multi-source data set, and inputs these new data into the existing deep learning model for forward propagation to calculate the error between the predicted result and the actual value. Based on this error, the photovoltaic load forecasting system can use the backpropagation algorithm to fine-tune the weights and biases of the model, so that the existing deep learning model can continue to adapt to the dynamic changes in the photovoltaic power generation environment, such as seasonal changes, changes in equipment performance, changes in electricity consumption patterns, etc., so as to maintain the accuracy and real-time nature of the prediction.

[0063] The following is an introduction to the model construction process. Specifically, after the photovoltaic load forecasting system constructs a multi-source data set, it can train the deep learning model in any of two ways. Specifically, the first solution can be seen in steps S203 and S204, first splitting the multi-source data set according to the characteristics of each deep learning model to obtain a dedicated training set corresponding to each deep learning model for training; the second solution can be seen in steps S205 and S206, directly using the pre-processed multi-source data set to train each deep learning model.

[0064] Through the above embodiments, the photovoltaic load forecasting system continuously acquires new data generated during the real-time operation of the photovoltaic power generation system, and regularly performs incremental learning on the trained deep learning model, so that the model can continuously track the changing trend of photovoltaic load characteristics, adaptively optimize and update the forecasting model, and enhance the dynamic adaptability and long-term stability of the photovoltaic load forecasting results.

[0065] S203. Divide the multi-source data set according to the preset processing data type of each deep learning model to obtain a dedicated training set corresponding to each deep learning model.

[0066] The photovoltaic load forecasting system divides the multi-source data set in a targeted manner according to the characteristics of different deep learning models and the preset processing data types, and creates a dedicated training set for each model, so as to give full play to the advantages of each model and improve the accuracy and efficiency of the overall forecast.

[0067] Specifically, the photovoltaic load forecasting system determines the data types that are most suitable for processing based on the structure and characteristics of each deep learning model. For example, long short-term memory networks (LSTM) are generally good at processing time series data, while convolutional neural networks (CNN) are more suitable for processing spatially related data. Based on these characteristics, the photovoltaic load forecasting system defines a data selection strategy for each model.

[0068] Then, according to these strategies, the corresponding data subsets (i.e., dedicated training sets) are screened and extracted from the multi-source data sets. For example, for the LSTM model, the photovoltaic load forecasting system may select continuous time series data; for the CNN model, data containing spatial information may be selected, such as photovoltaic panel output data in different geographical locations. Of course, the user can also directly set the data type that each deep learning model is most suitable for processing, and then directly determine the data subset (i.e., dedicated training set) corresponding to each model, which is not limited here.

[0069] S204: Use a dedicated training set to train the deep learning model to obtain multiple different target deep learning models.

[0070] After the photovoltaic load forecasting system obtains the dedicated training set corresponding to each deep learning model in step S203, the dedicated training set is used to train the corresponding deep learning model. The specific training process is basically the same as the use of multi-source data sets to train multiple different deep learning models in step S102, and will not be repeated here.

[0071] Through the above embodiments, in order to enable each model to play its strengths, the photovoltaic load forecasting system selects the data subset that each model is best at processing from the multi-source data set as its exclusive training set. Each trained model can make reliable predictions in its respective areas of expertise, thereby improving the accuracy and reliability of photovoltaic load forecasting.

[0072] S205. Preprocess the multi-source data set according to the training requirements of each deep learning model to obtain multiple feature training sets.

[0073] Different deep learning models have different requirements for the format, features and quality of input data. Therefore, before model training, the photovoltaic load forecasting system will appropriately preprocess the original multi-source data set so that it can best meet the needs of each model.

[0074] S206: Use the feature training set to train the deep learning model to obtain multiple different target deep learning models.

[0075] After data preprocessing, the photovoltaic load forecasting system obtains feature training sets optimized for different deep learning models. Then, the photovoltaic load forecasting system uses these feature training sets to train the corresponding deep learning models respectively, and obtains multiple target deep learning models with photovoltaic load forecasting capabilities.

[0076] Through the above embodiments, different deep learning models have different requirements for the format and quality of training data. In order to give full play to the modeling capabilities of each model, the photovoltaic load forecasting system performs customized preprocessing on the original multi-source data set according to the characteristics of each deep learning model, and extracts feature training sets that can match each model interface respectively. It not only takes into account the needs of multi-model integration strategies for data diversity, but also takes into account the personalized requirements of different models for data quality. While improving the prediction performance of a single model, it also lays the foundation for subsequent model fusion.

[0077] S207. Input the multi-source data acquired in real time into different target deep learning models respectively, and output the photovoltaic load prediction result corresponding to each target deep learning model.

[0078] This step is the same as step S104 and will not be repeated here.

[0079] S208. Determine the corresponding weighted value of each target deep learning model based on its accuracy on the validation set.

[0080] Since the photovoltaic load forecasting system adopts a multi-model integration strategy, it is necessary to reasonably weigh the prediction results of each target deep learning model so that they can maximize their respective modeling advantages.

[0081] Specifically, after completing the training of multiple different deep learning models, the photovoltaic load forecasting system will reserve a portion of historical data as a validation set to evaluate the generalization ability and prediction accuracy of each model. For example, the system can select the data of the last three months as a validation set, input it into different target deep learning models for inference, obtain a series of photovoltaic load forecast values ​​corresponding to each model, and then compare these forecast values ​​with the actual measured values ​​of the same period, and calculate the scoring indicators of each model on the validation set, such as root mean square error (RMSE), determination coefficient (R 2 ) etc., not limited here.

[0082] Next, the photovoltaic load forecasting system normalizes the scores of each model on the validation set to obtain a set of weight values ​​as the basic weighting for model fusion, where the weight value of each model is proportional to the scoring index.

[0083] For example, assuming that there are three target deep learning models A, B, and C in the photovoltaic load forecasting system, and their RMSE scores on the validation set are 10, 12, and 15 respectively, then their corresponding weighted values ​​can be set to A: 0.4, B: 0.33, and C: 0.27. In this way, the photovoltaic load forecasting system can effectively utilize the different advantages of each model, making the fused photovoltaic load forecasting results more accurate and reliable.

[0084] Through the above embodiment, the photovoltaic load forecasting system evaluates the contribution of each model to the final forecast result based on the forecast accuracy of the model on the validation set, and the model with higher accuracy is given a greater weight. This method can make fuller use of the forecast results of models with good performance and weaken the influence of poor models, so that the forecast results of multi-model fusion are as close to the true value as possible.

[0085] S209: Determine the final photovoltaic load prediction result based on the photovoltaic load prediction result and the weighted value.

[0086] In the above steps, the photovoltaic load forecasting system has obtained the corresponding photovoltaic load forecasting results from multiple different target deep learning models, and assigned a weighted value to each model based on the validation set. Then, the forecasting system merges these scattered forecasting results by weighted average to obtain the final photovoltaic load forecast value.

[0087] For example, assuming that at a certain prediction time point, the prediction value of model A is 100kW, the prediction value of model B is 120kW, and the prediction value of model C is 90kW, and their corresponding weighted values ​​are 0.4, 0.33, and 0.27, respectively, then the final photovoltaic load prediction value is: 100×0.4+120×0.33+90×0.27=104kW.

[0088] Of course, in addition to using weighted average to fuse the prediction results, the photovoltaic load forecasting system can also use other more complex model fusion strategies, such as the Bayesian model averaging (BMA)-based method, hierarchical weighted fusion, etc. The specific fusion strategy can be selected according to actual needs and is not limited here.

[0089] After generating the final photovoltaic load forecast result, the photovoltaic load forecast system will evaluate the forecast result to verify whether there is any abnormality. Figure 3 The figure is a schematic diagram of a process of evaluating the final photovoltaic load prediction result by the photovoltaic load prediction system in the embodiment of the present application.

[0090] S301. Obtain an actual measured value of the photovoltaic load.

[0091] The photovoltaic load prediction system will record the actual load data of the photovoltaic power generation system at each detection time point, and match the actual load data with the final photovoltaic load prediction result at the corresponding detection time point, so as to conduct error assessment on the final photovoltaic load prediction result based on the difference between the two.

[0092] S202: Generate an error curve for photovoltaic load prediction based on the difference between the actual measurement value and the final photovoltaic load prediction result.

[0093] Specifically, the photovoltaic load prediction system calculates the difference between the actual measurement value and the final photovoltaic load prediction result, and plots the change of the difference over time in the form of a curve, that is, forming an error curve of the photovoltaic load prediction, thereby showing the user the degree of deviation of the prediction result.

[0094] Through the above embodiments, the photovoltaic load prediction system introduces real photovoltaic load measurement values ​​as evaluation criteria, quantifies the deviation between the multi-model fusion prediction results and objective facts, and visually displays the dynamic change trend of the prediction error over time through the error curve, so as to facilitate users to evaluate the accuracy of the final photovoltaic load prediction results.

[0095] S203: Count the number of final photovoltaic load prediction results whose difference exceeds a preset range threshold within a preset time period.

[0096] In this step, the user can set a preset time period (such as 15 days) and a preset range threshold (such as 20KW) in the photovoltaic load forecasting system. When the difference at a certain detection time point exceeds the preset range threshold, it is determined that the final photovoltaic load forecast result at that time point is abnormal.

[0097] Furthermore, the photovoltaic load forecasting system counts the number of differences exceeding a preset range threshold within the preset time period, and evaluates the stability and accuracy of the models involved in the forecasting by counting the number of abnormal final photovoltaic load forecasting results.

[0098] S204: If it is detected that the number of final photovoltaic load prediction results exceeds a preset number threshold, it is determined that the final photovoltaic load prediction result is abnormal.

[0099] For example, in one embodiment, the preset number threshold set by the technician based on actual needs is 3. If 15 photovoltaic load tests are performed within the preset time period (15 days), on this basis, it is detected that there are 4 test results in which the difference between the actual measurement value and the final photovoltaic load prediction result exceeds the preset range threshold, then it is determined that the final photovoltaic load prediction result within the preset time period is abnormal, and a prediction abnormality warning is sent to the technician.

[0100] The photovoltaic load prediction system according to the embodiment of the present invention is applied to electronic equipment. Figure 4 A schematic diagram of the architecture of an electronic device suitable for implementing an embodiment of the present invention is shown.

[0101] It should be noted that Figure 4 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0102] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructions (computer programs), or by controlling related hardware through instructions (computer programs), and the instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. The electronic device of this embodiment includes a storage medium and a processor, wherein a plurality of instructions are stored in the storage medium, and the instructions can be loaded by the processor to execute any step of the method provided in the embodiment of the present invention.

[0103] Specifically, the storage medium and the processor are electrically connected directly or indirectly to realize data transmission or interaction. For example, these elements can be electrically connected to each other through one or more signal lines. The storage medium stores computer execution instructions for implementing the data access control method, including at least one software function module that can be stored in the storage medium in the form of software or firmware. The processor executes various functional applications and data processing by running the software program and module stored in the storage medium. The storage medium can be, but is not limited to, random access storage medium (Random Access Memory, referred to as: RAM), read-only storage medium (Read Only Memory, referred to as: ROM), programmable read-only storage medium (Programmable Read-Only Memory, referred to as: PROM), erasable read-only storage medium (Erasable Programmable Read-Only Memory, referred to as: EPROM), electrically erasable read-only storage medium (Electric Erasable Programmable Read-Only Memory, referred to as: EEPROM), etc. Among them, the storage medium is used to store programs, and the processor executes the program after receiving the execution instruction.

[0104] Furthermore, the software programs and modules in the above-mentioned storage medium may also include an operating system, which may include various software components and / or drivers for managing system tasks (such as memory management, storage device control, power management, etc.), and may communicate with various hardware or software components to provide an operating environment for other software components. The processor may be an integrated circuit chip having signal processing capabilities. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc., which may implement or execute the various methods, steps, and logic flow diagrams disclosed in this embodiment. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0105] Since the instructions stored in the storage medium can execute the steps in any method provided in the embodiments of the present invention, the beneficial effects of any method provided in the embodiments of the present invention can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0106] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A photovoltaic load forecasting method based on multi-source data deep learning, applied to photovoltaic load forecasting system, characterized in that: The method comprises: Acquire historical interference data and construct a multi-source data set, wherein the historical interference data includes power generation data, power generation efficiency data, transmission loss data, historical weather data, and historical load data of the photovoltaic power generation system; Using the multi-source data sets to train multiple different deep learning models respectively, to obtain multiple different target deep learning models; Get the weighted value corresponding to each target deep learning model; Inputting the multi-source data acquired in real time into the different target deep learning models respectively, and outputting the photovoltaic load prediction result corresponding to each target deep learning model, wherein the multi-source data corresponds to the historical interference data one by one; The final photovoltaic load prediction result is calculated based on the photovoltaic load prediction result output by each target deep learning model and the corresponding weighted value.

2. The method according to claim 1, characterized in that: After the step of using the multi-source data sets to respectively train a plurality of different deep learning models to obtain a plurality of different target deep learning models, the method further includes: Acquire new multi-source data in real time; updating the multi-source data set according to the new multi-source data; Regularly use the newly added data in the multi-source dataset to perform incremental training on the target deep learning model.

3. The method according to claim 1, characterized in that: The step of obtaining the weighted value corresponding to each target deep learning model specifically includes: Get the accuracy of each target deep learning model on the validation set; The weighted value corresponding to each target deep learning model is determined according to the ratio between the accuracies.

4. The method according to claim 1, characterized in that: The step of using the multi-source data sets to respectively train a plurality of different deep learning models to obtain a plurality of different target deep learning models specifically includes: According to the training requirements of each deep learning model, the multi-source data set is preprocessed to obtain multiple feature training sets, and the feature training sets correspond to the deep learning models one by one; The deep learning model is trained using the feature training set to obtain multiple different target deep learning models.

5. The method according to claim 1, characterized in that The step of using the multi-source data sets to respectively train a plurality of different deep learning models to obtain a plurality of different target deep learning models specifically includes: Dividing the multi-source data set according to the processing data type preset by each deep learning model to obtain a dedicated training set corresponding to each deep learning model, wherein the dedicated training set is a subset of the multi-source data set; The deep learning model is trained using a dedicated training set to obtain multiple different target deep learning models.

6. The method according to claim 1, characterized in that After the step of calculating the final photovoltaic load prediction result according to the photovoltaic load prediction result output by each target deep learning model and the corresponding weighted value, the method further includes: Get actual measurements of PV load; Generating an error curve for photovoltaic load prediction according to the difference between the actual measurement value and the final photovoltaic load prediction result; The final photovoltaic load prediction result is evaluated according to the fluctuation condition of the error curve.

7. The method according to claim 6, characterized in that The step of evaluating the final photovoltaic load prediction result according to the fluctuation condition of the error curve specifically includes: Counting the number of final photovoltaic load prediction results whose difference exceeds a preset range threshold within a preset time period; If it is detected that the number of the final photovoltaic load prediction results exceeds a preset number threshold, it is determined that the final photovoltaic load prediction result is abnormal.

8. A photovoltaic load forecasting system, characterized in that: The photovoltaic load prediction system includes: one or more processors and memory; The memory is coupled to the one or more processors, and the memory is used to store computer program codes, wherein the computer program codes include computer instructions, and the one or more processors call the computer instructions to enable the photovoltaic load forecasting system to perform the method according to any one of claims 1 to 7.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instruction is executed on a photovoltaic load forecasting system, the photovoltaic load forecasting system is caused to execute the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that When the computer program product runs on a photovoltaic load forecasting system, the photovoltaic load forecasting system is enabled to perform the method according to any one of claims 1 to 7.

Citation Information

Cited By

  • Energy power generation control method driven by multi-source power coupling model

    CN119382247A

  • Energy generation control method driven by multi-source power coupling model

    CN119382247B

  • Flange connection tightness calculation modeling method, calculation method, system, equipment and medium

    CN119885497A

  • Flange Connection Tightness Calculation Modeling Method, Calculation Method, System, Equipment and Medium

    CN119885497B

  • New energy output evaluation system based on artificial intelligence meteorological large model

    CN120181636A